ALASKA STATE LEGISLATURE  HOUSE JUDICIARY STANDING COMMITTEE  February 28, 2025 1:01 p.m. DRAFT MEMBERS PRESENT Representative Andrew Gray, Chair Representative Chuck Kopp, Vice Chair Representative Ted Eischeid Representative Genevieve Mina Representative Mia Costello Representative Jubilee Underwood MEMBERS ABSENT  Representative Sarah Vance COMMITTEE CALENDAR  PRESENTATION(S): LEGAL AND ETHICAL IMPLICATIONS OF ARTIFICIAL INTELLIGENCE - HEARD PREVIOUS COMMITTEE ACTION  No previous action to record WITNESS REGISTER GUARAV KHANA, PhD, Senior Manager of Data Science and Digital Journeys Cisco Systems; AI Leadership Instructor, Stanford University Juneau, Alaska POSITION STATEMENT: Presented on AI foundations during the Legal and Ethical Implications of Artificial Intelligence presentation. ROSE FELICIANO, Executive Director for Washington and Northwest TechNet Juneau, Alaska POSITION STATEMENT: Presented on AI policy during the Legal and Ethical Implications of Artificial Intelligence presentation. ACTION NARRATIVE 1:01:03 PM CHAIR ANDREW GRAY called the House Judiciary Standing Committee meeting to order at 1:01 p.m. Representatives Costello, Mina, Eischeid, and Gray were present at the call to order. Representatives Underwood and Kopp arrived as the meeting was in progress. ^PRESENTATION(S): Legal and Ethical Implications of Artificial Intelligence PRESENTATION(S): Legal and Ethical Implications of Artificial  Intelligence    1:01:47 PM CHAIR GRAY announced that the only order of business would be the Legal and Ethical Implications of Artificial Intelligence presentation. 1:02:27 PM GUARAV KHANA, PhD, Senior Manager of Data Science and Digital Journeys, Cisco Systems; AI Leadership Instructor, Stanford University, gave a PowerPoint presentation, titled "AI Foundations," [hard copy included in the committee packet]. Beginning on slide 2, he recalled the historic launch of ChatGPT on November 30, 2022. On slide 3, he elaborated on the economic impacts of AI, which is estimated to add between $2.6 trillion to $4.4 trillion annually. Continuing to slide 4, he explained that AI is fundamentally good at detecting patterns and anomalies across large datasets. On slide 5, he likened AI to a rocket ship with a quote from Andrew Ng. He outlined the journey of machine learning - or AI - on slide 6, recalling that it experienced a renaissance in the late 2000s with tasks like detecting fraudulent credit card transactions. Its objectives grew more sophisticated with the implementation of object identification and x-ray analysis in the 2010s, growing to self- driving cars and text/video generation at present. On slide 7, he spoke to AI's forecasted contribution to the global economy of $15.7 trillion by 2030, noting that thus far, AI has always exceeded projections. Slide 8 highlighted use cases of AI in state government for things like court filings, traffic decongestion, and chatbots. Slides 9 through 11 gave examples of how ChatGPT is being used. On slide 12, he explained the transformative power of AI. He shared an example of how generative AI systems make errors to illustrated the importance of input and phrasing and why it matters. Slide 14 outlined the latest research and innovation on grounding, alignment, and safety. Grounding refers to a large language model's (LLM) ability to provide accurate, reliable, and verifiable answers. Alignment refers to the ability avoid harmful, biased, or inappropriate outputs while remaining useful; the ability to match human expectations and societal/organization norms. The concept of safety refers to the ability for bad actors to compromise these systems and output harmful and toxic content. 1:24:24 PM DR. KHANA, in response to a series of questions, said the difference between alignment and safety is the ability for the LLM to sound like one's brand versus defense against outputting toxic language; guardrails are built into the modes to prevent safety violations and toxic speech; AI has no concept of reasoning and instead, gives the illusion of thinking like a human by predicting words; there is a meaning, depth, and emotion expressed by humans that is not present in these systems. 1:35:01 PM DR. KHANA resumed the presentation on slides 15-17 with examples of how to "hack" an LLM for a secret password. He explained that AI can be tricked into revealing something it shouldn't or generating toxic content. He concluded that these systems are wonderful and innovative, but their flaws must be acknowledged and mitigated. 1:38:52 PM ROSE FELICIANO, Executive Director for Washington and Northwest, TechNet, gave a presentation on TechNet and AI legislation. She shared several policy principles that TechNet urges legislatures to consider when crafting AI related legislation. First, policy makers should avoid blanket prohibitions on AI, machine learning, or other forms of automated decision making and instead reserve any restrictions for specific use cases that present clear, demonstrated risk of unacceptable harm. Second, policy makers should leverage existing authorities under state law that provide substantial anti-discrimination and civil rights protections and limit new authorities specific to the operation of AI where existing authorities are inadequate. Third, policy makers should ensure any requirements on automated decision tools focus on high risk uses where decisions are based solely on the automated decisions. They should avoid labeling entire sectors as inherently high risk and focus on specific outcomes that involve the loss of life or liberty or have significant legal effects on people. Fourth, interoperability is a huge concern for the technology industry. She urged legislators to rely on established national and international framework as a guide for developing policy to ensure operability and avoid patchwork. She encouraged members to allow measures taken to comply with one law or regulation to satisfy the requirements for another applicable law or regulation. She shared that last year, TechNet had tracked 476 AI bills across the country that largely focused on the creation of AI task forces, election misinformation, protections against child sexual abuse material (CSAM), and safeguarding against potential bias in automated decision tools. CHAIR GRAY acknowledged that protecting the public from harm may cost money and may require restrictions on the technology industry. He emphasized that the committee's goal is public protection, not the protection of a particular company. 1:48:34 PM MS. FELICIANO shared an example involving Zillow and the use of artificial staging for real estate listings and how that might feel deceitful to buyers. In response to a series of questions about the use of AI in election material, she said a number of states have passed AI disclosure laws. TechNet believes that it's the candidate's responsibility [to represent themselves honestly], and that this issue is much more important if AI is being used harmfully or to disparage another candidate, as opposed to being used to enhance one's physical appearance, for example. DR. KHANA, in response to a question from the chair, indicated that it would be possible to quantify the use of AI in a particular advertisement or video so that a usage threshold could be implemented for disclosure, but implementing uniform standards would be difficult because the variety of instances is so vast. He suggested that if AI was used to craft something entirely, a watermark could be used. Further, he shared his belief that some of these standards would be determined in court. He reasoned that disclosing a modicum use of AI would be unreasonable. MS. FELICIANO added that the state of Washington is considering a bill that would require watermarking, also referred to as "content provenance," for content altered by AI. 1:57:41 PM MS. FELICIANO, in response to a series of questions about data privacy, reported that 19 states had adopted data privacy laws and suggested that all states should have one as a starting point. With regard to health data, she said it must follow Health Insurance Portability and Accountability Act (HIPAA) guidelines. In response to a question about collaboration with law enforcement, she said many AI companies have dedicated staff whose job it is to look for CSAM and report it to the National Center for Missing and Exploited Children, per federal law. In response to a question about the regulation of data centers, she said it's important that electric utilities and regional planning companies are anticipating future need to prepare for environmental and energy impacts. 2:06:36 PM DR. KHANA agreed that an updated data privacy policy would be a good starting point. He explained that the heart of a good data privacy policy includes categorizing data and creating a taxonomy of confidential versus public data or allowable versus restricted data, for example. He said there's a way to do data loss protection as well. Many newer generative AI models are not that energy intensive and can be built for energy efficiency. He added that smaller models can outperform larger models for certain companies if they're trained to do specific tasks well, versus trained to do everything like ChatGPT. He opined that there wouldn't be an exponential need for energy that crashes the grid; however, the need for power has been underestimated. In response to a series of questions, he reported that it takes months to build large generative AI chatbots and days for smaller models, because they are often derivatives of the larger models; guardrails are not required by law, which could be mandated for any models that would be used by the state along with other rules and layers of protection. 2:12:48 PM MS. FELICIANO offered to follow up on two questions about whether the Uniform Law Commission had standardized statutory language and which states Alaska could look to for crafting policy that imposes guardrails on new generative AI chatbots. DR. KHANA spoke to the importance of striking a balance between generalized and overly strict language to avoid the need to revisit the policy as technology evolves. Be recommended erring on the side of caution and including more rather than less to be as comprehensive as possible. 2:17:19 PM ADJOURNMENT  There being no further business before the committee, the House Judiciary Standing Committee meeting was adjourned at 2:17 p.m.